TimesFM GAT Model for Industrial Recipe Anomaly Attribution with Batch Process Logs and Quality Labels

Batch manufacturing systems save detailed event logs, phase records, set-point traces and delayed quality labels, but abnormal quality changes are often only identified by isolated control charts or operator experience. This paper proposes a TimesFM-GAT model for recipe anomaly attribution in batch processes, and aligns pretrained temporal representations with a recipe graph that connects units, phases, sensors, manipulated variables, and laboratory quality results. TimesFM is used to encode multiscale log windows, a graph attention network propagates contextual evidence, and quality-supervised contrastive learning constrains attribution. A synthetic but process-faithful fermentation and specialty-chemical benchmark with 18,240 batches, 42 process variables, 11 recipe phases and four quality labels were built to evaluate detection, localization and diagnostic stability. Compared with the LSTM, TCN, TimesFM-only and GAT-only baselines, the proposed model increased the batch-level F1 score from 0.842 to 0.913, improved phase localization accuracy from 0.781 to 0.864, and reduced mean attribution drift by 27.6% under sensor missingness and recipe changeover conditions. Based on the above results, the foundation-model temporal embeddings and process-graph reasoning are complementary in identifying quality-related abnormal evidence in logged batch operations. The proposed framework offers a feasible path for recipe monitoring systems that require early warning, interpretable root-cause ranking and transferable deployment across related product grades.